Inversion of nitrogen content in apple leaves using an explainable PSO-CNN model

Introduction: The integration of hyperspectral technology with machine learning and deep learning algorithms offers an effective method for accurately and non-destructively estimating the percentage of nitrogen in apple tree leaves, as well as for rapid nutrient diagnosis. Methods: This study was conducted in apple orchards in Qixia City, Shandong Province, where hyperspectral data were collected from Red Fuji apple trees during the new-shoot-stop-growing stage (NSS) and the autumn-shoot-stop-growing stage (ASS). Following hyperspectral preprocessing and characteristic wavelength selection, regression models-including random forest (RF), support vector machine (SVM), convolutional neural network (CNN), and particle swarm optimization convolutional neural network (PSO-CNN)-were developed and compared. Results: = 1.529), with uniformly distributed errors and no systematic bias, indicating good generalization capability and stability. SHAP analysis revealed that the PSO-CNN model primarily relied on the near-infrared and short-wave infrared bands, where reflectance was negatively correlated with nitrogen content. Discussion: These spectral regions are closely associated with leaf biochemical structure, suggesting that the model predicts nitrogen content by capturing spectral information related to leaf biochemical characteristics. Overall, the PSO-CNN model improves nitrogen prediction performance by expanding the hyperparameter search space while preserving the CNN architecture, enabling rapid and accurate nutrient diagnosis in apple leaves.